Weakly-supervised Disentangling with Recurrent Transformations for 3D View Synthesis

less than 1 minute read

Published:

observation

  • human take time proportional to rotated angles in matching images of the same object
  • recurrent: rotate 15 degrees once in a step

model

  • encoder-decoder
  • hidden unit: identity unit + pose unit
    • controlled by action unit
    • mimic the underlying manifold step by step
    • identity unit being shared while pose unit keeps changing
  • also predict mask